Skip to content

Instantly share code, notes, and snippets.

@justinruggles

justinruggles/gist:3213077

Created July 31, 2012 02:50
Show Gist options

  • Select an option

    Learn more about clone URLs
  • Save justinruggles/3213077 to your computer and use it in GitHub Desktop.

Select an option

Learn more about clone URLs
Save justinruggles/3213077 to your computer and use it in GitHub Desktop.
Download ZIP
Raw
gistfile1.diff
From 03f42e9e885ce8e5f29e3f04e2c98e5352f69df4 Mon Sep 17 00:00:00 2001
From: "Ronald S. Bultje" <rsbultje@gmail.com>
Date: Mon, 30 Jul 2012 13:00:04 -0700
Subject: [PATCH] fft: port FFT and IMDCT functions for 3dnow to yasm.
---
libavcodec/x86/Makefile | 2 -
libavcodec/x86/fft_3dn.c | 23 ------
libavcodec/x86/fft_3dn2.c | 173 --------------------------------------------
libavcodec/x86/fft_mmx.asm | 169 +++++++++++++++++++++++++++++++++++++++----
4 files changed, 153 insertions(+), 214 deletions(-)
delete mode 100644 libavcodec/x86/fft_3dn.c
delete mode 100644 libavcodec/x86/fft_3dn2.c
diff --git a/libavcodec/x86/Makefile b/libavcodec/x86/Makefile
index 1d26357..0ae70b2 100644
--- a/libavcodec/x86/Makefile
+++ b/libavcodec/x86/Makefile
@@ -38,8 +38,6 @@ YASM-OBJS-$(CONFIG_AAC_DECODER) += x86/sbrdsp.o
YASM-OBJS-$(CONFIG_AC3DSP) += x86/ac3dsp.o
YASM-OBJS-$(CONFIG_DCT) += x86/dct32_sse.o
YASM-OBJS-$(CONFIG_ENCODERS) += x86/dsputilenc_yasm.o
-YASM-OBJS-FFT-$(HAVE_AMD3DNOW) += x86/fft_3dn.o
-YASM-OBJS-FFT-$(HAVE_AMD3DNOWEXT) += x86/fft_3dn2.o
YASM-OBJS-$(CONFIG_FFT) += x86/fft_mmx.o \
$(YASM-OBJS-FFT-yes)
YASM-OBJS-$(CONFIG_H264CHROMA) += x86/h264_chromamc.o \
diff --git a/libavcodec/x86/fft_3dn.c b/libavcodec/x86/fft_3dn.c
deleted file mode 100644
index 5a4d3ad..0000000
--- a/libavcodec/x86/fft_3dn.c
+++ /dev/null
@@ -1,23 +0,0 @@
-/*
- * FFT/MDCT transform with 3DNow! optimizations
- * Copyright (c) 2008 Loren Merritt
- *
- * This file is part of Libav.
- *
- * Libav is free software; you can redistribute it and/or
- * modify it under the terms of the GNU Lesser General Public
- * License as published by the Free Software Foundation; either
- * version 2.1 of the License, or (at your option) any later version.
- *
- * Libav is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
- * Lesser General Public License for more details.
- *
- * You should have received a copy of the GNU Lesser General Public
- * License along with Libav; if not, write to the Free Software
- * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
- */
-
-#define EMULATE_3DNOWEXT
-#include "fft_3dn2.c"
diff --git a/libavcodec/x86/fft_3dn2.c b/libavcodec/x86/fft_3dn2.c
deleted file mode 100644
index e684cc7..0000000
--- a/libavcodec/x86/fft_3dn2.c
+++ /dev/null
@@ -1,173 +0,0 @@
-/*
- * FFT/MDCT transform with Extended 3DNow! optimizations
- * Copyright (c) 2006-2008 Zuxy MENG Jie, Loren Merritt
- *
- * This file is part of Libav.
- *
- * Libav is free software; you can redistribute it and/or
- * modify it under the terms of the GNU Lesser General Public
- * License as published by the Free Software Foundation; either
- * version 2.1 of the License, or (at your option) any later version.
- *
- * Libav is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
- * Lesser General Public License for more details.
- *
- * You should have received a copy of the GNU Lesser General Public
- * License along with Libav; if not, write to the Free Software
- * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
- */
-
-#include "libavutil/x86_cpu.h"
-#include "libavcodec/dsputil.h"
-#include "fft.h"
-
-DECLARE_ALIGNED(8, static const unsigned int, m1m1)[2] = { 1U<<31, 1U<<31 };
-
-#ifdef EMULATE_3DNOWEXT
-#define PSWAPD(s,d)\
- "movq "#s","#d"\n"\
- "psrlq $32,"#d"\n"\
- "punpckldq "#s","#d"\n"
-#define ff_fft_calc_3dnow2 ff_fft_calc_3dnow
-#define ff_fft_dispatch_3dnow2 ff_fft_dispatch_3dnow
-#define ff_fft_dispatch_interleave_3dnow2 ff_fft_dispatch_interleave_3dnow
-#define ff_imdct_calc_3dnow2 ff_imdct_calc_3dnow
-#define ff_imdct_half_3dnow2 ff_imdct_half_3dnow
-#else
-#define PSWAPD(s,d) "pswapd "#s","#d"\n"
-#endif
-
-void ff_fft_dispatch_3dnow2(FFTComplex *z, int nbits);
-void ff_fft_dispatch_interleave_3dnow2(FFTComplex *z, int nbits);
-
-void ff_fft_calc_3dnow2(FFTContext *s, FFTComplex *z)
-{
- int n = 1<<s->nbits;
- int i;
- ff_fft_dispatch_interleave_3dnow2(z, s->nbits);
- __asm__ volatile("femms");
- if(n <= 8)
- for(i=0; i<n; i+=2)
- FFSWAP(FFTSample, z[i].im, z[i+1].re);
-}
-
-void ff_imdct_half_3dnow2(FFTContext *s, FFTSample *output, const FFTSample *input)
-{
- x86_reg j, k;
- long n = s->mdct_size;
- long n2 = n >> 1;
- long n4 = n >> 2;
- long n8 = n >> 3;
- const uint16_t *revtab = s->revtab;
- const FFTSample *tcos = s->tcos;
- const FFTSample *tsin = s->tsin;
- const FFTSample *in1, *in2;
- FFTComplex *z = (FFTComplex *)output;
-
- /* pre rotation */
- in1 = input;
- in2 = input + n2 - 1;
-#ifdef EMULATE_3DNOWEXT
- __asm__ volatile("movd %0, %%mm7" ::"r"(1U<<31));
-#endif
- for(k = 0; k < n4; k++) {
- // FIXME a single block is faster, but gcc 2.95 and 3.4.x on 32bit can't compile it
- __asm__ volatile(
- "movd %0, %%mm0 \n"
- "movd %2, %%mm1 \n"
- "punpckldq %1, %%mm0 \n"
- "punpckldq %3, %%mm1 \n"
- "movq %%mm0, %%mm2 \n"
- PSWAPD( %%mm1, %%mm3 )
- "pfmul %%mm1, %%mm0 \n"
- "pfmul %%mm3, %%mm2 \n"
-#ifdef EMULATE_3DNOWEXT
- "movq %%mm0, %%mm1 \n"
- "punpckhdq %%mm2, %%mm0 \n"
- "punpckldq %%mm2, %%mm1 \n"
- "pxor %%mm7, %%mm0 \n"
- "pfadd %%mm1, %%mm0 \n"
-#else
- "pfpnacc %%mm2, %%mm0 \n"
-#endif
- ::"m"(in2[-2*k]), "m"(in1[2*k]),
- "m"(tcos[k]), "m"(tsin[k])
- );
- __asm__ volatile(
- "movq %%mm0, %0 \n\t"
- :"=m"(z[revtab[k]])
- );
- }
-
- ff_fft_dispatch_3dnow2(z, s->nbits);
-
-#define CMUL(j,mm0,mm1)\
- "movq (%2,"#j",2), %%mm6 \n"\
- "movq 8(%2,"#j",2), "#mm0"\n"\
- "movq %%mm6, "#mm1"\n"\
- "movq "#mm0",%%mm7 \n"\
- "pfmul (%3,"#j"), %%mm6 \n"\
- "pfmul (%4,"#j"), "#mm0"\n"\
- "pfmul (%4,"#j"), "#mm1"\n"\
- "pfmul (%3,"#j"), %%mm7 \n"\
- "pfsub %%mm6, "#mm0"\n"\
- "pfadd %%mm7, "#mm1"\n"
-
- /* post rotation */
- j = -n2;
- k = n2-8;
- __asm__ volatile(
- "1: \n"
- CMUL(%0, %%mm0, %%mm1)
- CMUL(%1, %%mm2, %%mm3)
- "movd %%mm0, (%2,%0,2) \n"
- "movd %%mm1,12(%2,%1,2) \n"
- "movd %%mm2, (%2,%1,2) \n"
- "movd %%mm3,12(%2,%0,2) \n"
- "psrlq $32, %%mm0 \n"
- "psrlq $32, %%mm1 \n"
- "psrlq $32, %%mm2 \n"
- "psrlq $32, %%mm3 \n"
- "movd %%mm0, 8(%2,%0,2) \n"
- "movd %%mm1, 4(%2,%1,2) \n"
- "movd %%mm2, 8(%2,%1,2) \n"
- "movd %%mm3, 4(%2,%0,2) \n"
- "sub $8, %1 \n"
- "add $8, %0 \n"
- "jl 1b \n"
- :"+r"(j), "+r"(k)
- :"r"(z+n8), "r"(tcos+n8), "r"(tsin+n8)
- :"memory"
- );
- __asm__ volatile("femms");
-}
-
-void ff_imdct_calc_3dnow2(FFTContext *s, FFTSample *output, const FFTSample *input)
-{
- x86_reg j, k;
- long n = s->mdct_size;
- long n4 = n >> 2;
-
- ff_imdct_half_3dnow2(s, output+n4, input);
-
- j = -n;
- k = n-8;
- __asm__ volatile(
- "movq %4, %%mm7 \n"
- "1: \n"
- PSWAPD((%2,%1), %%mm0)
- PSWAPD((%3,%0), %%mm1)
- "pxor %%mm7, %%mm0 \n"
- "movq %%mm1, (%3,%1) \n"
- "movq %%mm0, (%2,%0) \n"
- "sub $8, %1 \n"
- "add $8, %0 \n"
- "jl 1b \n"
- :"+r"(j), "+r"(k)
- :"r"(output+n4), "r"(output+n4*3),
- "m"(*m1m1)
- );
- __asm__ volatile("femms");
-}
diff --git a/libavcodec/x86/fft_mmx.asm b/libavcodec/x86/fft_mmx.asm
index 527e215..125bbd8 100644
--- a/libavcodec/x86/fft_mmx.asm
+++ b/libavcodec/x86/fft_mmx.asm
@@ -396,7 +396,6 @@ fft32_interleave_avx:
%endif
INIT_XMM sse
-%define movdqa movaps
align 16
fft4_avx:
@@ -588,6 +587,7 @@ INIT_XMM sse
DECL_PASS pass_sse, PASS_BIG 1
DECL_PASS pass_interleave_sse, PASS_BIG 0
+%macro FFT_CALC_FUNC 0
cglobal fft_calc, 2,5,8
mov r3d, [r0 + FFTContext.nbits]
PUSH r1
@@ -597,23 +597,39 @@ cglobal fft_calc, 2,5,8
FFT_DISPATCH _interleave %+ SUFFIX, r1
POP rcx
POP r4
- cmp rcx, 4
+%if cpuflag(3dnow)
+ femms
+%endif
+ cmp rcx, 3+(mmsize/16)
jg .end
mov r2, -1
add rcx, 3
shl r2, cl
sub r4, r2
.loop
+%if mmsize == 8
+ pswapd m0, [r4 + r2 + 4]
+ mova [r4 + r2 + 4], m0
+%else
movaps xmm0, [r4 + r2]
movaps xmm1, xmm0
unpcklps xmm0, [r4 + r2 + 16]
unpckhps xmm1, [r4 + r2 + 16]
movaps [r4 + r2], xmm0
movaps [r4 + r2 + 16], xmm1
- add r2, 32
+%endif
+ add r2, mmsize*2
jl .loop
.end:
REP_RET
+%endmacro
+
+INIT_MMX 3dnow
+FFT_CALC_FUNC
+INIT_MMX 3dnow2
+FFT_CALC_FUNC
+INIT_XMM sse
+FFT_CALC_FUNC
cglobal fft_permute, 2,7,1
mov r4, [r0 + FFTContext.revtab]
@@ -648,6 +664,7 @@ cglobal fft_permute, 2,7,1
jl .loopcopy
REP_RET
+%macro IMDCT_CALC_FUNC 0
cglobal imdct_calc, 3,5,3
mov r3d, [r0 + FFTContext.mdctsize]
mov r4, [r0 + FFTContext.imdcthalf]
@@ -671,21 +688,50 @@ cglobal imdct_calc, 3,5,3
POP r3
lea r0, [r1 + 2*r3]
mov r2, r3
- sub r3, 16
+ sub r3, mmsize
neg r2
- movaps xmm2, [ps_m1m1m1m1]
+ mova m2, [ps_m1m1m1m1]
.loop:
- movaps xmm0, [r1 + r3]
- movaps xmm1, [r0 + r2]
- shufps xmm0, xmm0, 0x1b
- shufps xmm1, xmm1, 0x1b
- xorps xmm0, xmm2
- movaps [r0 + r3], xmm1
- movaps [r1 + r2], xmm0
- sub r3, 16
- add r2, 16
+%if mmsize == 8
+ PSWAPD m0, [r1 + r3]
+ PSWAPD m1, [r0 + r2]
+ pxor m0, m2
+%else
+ mova m0, [r1 + r3]
+ mova m1, [r0 + r2]
+ shufps m0, m0, 0x1b
+ shufps m1, m1, 0x1b
+ xorps m0, m2
+%endif
+ mova [r0 + r3], m1
+ mova [r1 + r2], m0
+ sub r3, mmsize
+ add r2, mmsize
jl .loop
+%if cpuflag(3dnow)
+ femms
+ RET
+%else
REP_RET
+%endif
+%endmacro
+
+%macro PSWAPD 2
+%if cpuflag(3dnow2)
+ pswapd %1, %2
+%else
+ mova %1, %2
+ psrlq %1, 32
+ punpckldq %1, %2
+%endif
+%endmacro
+
+INIT_MMX 3dnow
+IMDCT_CALC_FUNC
+INIT_MMX 3dnow2
+IMDCT_CALC_FUNC
+INIT_XMM sse
+IMDCT_CALC_FUNC
INIT_MMX 3dnow
%define mulps pfmul
@@ -775,6 +821,25 @@ INIT_XMM sse
%undef unpckhps
%macro PREROTATER 5 ;-2*k, 2*k, input+n4, tcos+n8, tsin+n8
+%if mmsize == 8 ; j+1-n4, n4-1-j, input+n4, tcos+n4, tsin+n4
+ movd m0, [%3+%2*4]
+ movd m1, [%4+%1*4-4]
+ punpckldq m0, [%3+%1*4-4]
+ punpckldq m1, [%5+%1*4-4]
+ mova m2, m0
+ PSWAPD m3, m1
+ pfmul m0, m1
+ pfmul m2, m3
+%if cpuflag(3dnow2)
+ pfpnacc m0, m2 ; FIXME 3dnow
+%else
+ mova m1, m0
+ punpckhdq m0, m2
+ punpckldq m1, m2
+ pxor m0, m7
+ pfadd m0, m1
+%endif
+%else
movaps xmm0, [%3+%2*4]
movaps xmm1, [%3+%1*4-0x10]
movaps xmm2, xmm0
@@ -795,6 +860,7 @@ INIT_XMM sse
movaps xmm0, xmm1
unpcklps xmm1, xmm2
unpckhps xmm0, xmm2
+%endif
%endmacro
%macro CMUL 6 ;j, xmm0, xmm1, 3, 4, 5
@@ -863,6 +929,40 @@ INIT_XMM sse
jl .post
%endmacro
+%macro CMUL_3DNOW 6
+ mova m6, [%1+%2*2]
+ mova %3, [%1+%2*2+8]
+ mova %4, m6
+ mova m7, %3
+ pfmul m6, [%5+%2]
+ pfmul %3, [%6+%2]
+ pfmul %4, [%6+%2]
+ pfmul m7, [%5+%2]
+ pfsub %3, m6
+ pfadd %4, m7
+%endmacro
+
+%macro POSROTATESHUF_3DNOW 5 ;j, k, z+n8, tcos+n8, tsin+n8
+.post:
+ CMUL_3DNOW %3, %1, m0, m1, %4, %5
+ CMUL_3DNOW %3, %2, m2, m3, %4, %5
+ movd [%3+%1*2+ 0], m0
+ movd [%3+%2*2+12], m1
+ movd [%3+%2*2+ 0], m2
+ movd [%3+%1*2+12], m3
+ psrlq m0, 32
+ psrlq m1, 32
+ psrlq m2, 32
+ psrlq m3, 32
+ movd [%3+%1*2+ 8], m0
+ movd [%3+%2*2+ 4], m1
+ movd [%3+%2*2+ 8], m2
+ movd [%3+%1*2+ 4], m3
+ sub %2, 8
+ add %1, 8
+ jl .post
+%endmacro
+
%macro DECL_IMDCT 1
cglobal imdct_half, 3,12,8; FFTContext *s, FFTSample *output, const FFTSample *input
%if ARCH_X86_64
@@ -876,37 +976,63 @@ cglobal imdct_half, 3,12,8; FFTContext *s, FFTSample *output, const FFTSample *i
%endif
mov r3d, [r0+FFTContext.mdctsize]
add r2, r3
+%if mmsize != 8
shr r3, 1
+%endif
mov rtcos, [r0+FFTContext.tcos]
mov rtsin, [r0+FFTContext.tsin]
add rtcos, r3
add rtsin, r3
+%if mmsize == 8
+ shr r3, 1
+%endif
%if ARCH_X86_64 == 0
push rtcos
push rtsin
%endif
+%if mmsize != 8
shr r3, 1
+%endif
mov rrevtab, [r0+FFTContext.revtab]
add rrevtab, r3
+%if mmsize == 8
+ shr r3, 1
+%endif
%if ARCH_X86_64 == 0
push rrevtab
%endif
+%if mmsize == 8
+ dec r3
+%else
sub r3, 4
-%if ARCH_X86_64
+%endif
+%if ARCH_X86_64 || mmsize == 8
xor r4, r4
sub r4, r3
%endif
+%if notcpuflag(3dnow2) && mmsize == 8
+ movd m7, [ps_m1m1m1m1]
+%endif
.pre:
%if ARCH_X86_64 == 0
;unspill
+%if mmsize != 8
xor r4, r4
sub r4, r3
- mov rtsin, [esp+4]
mov rtcos, [esp+8]
%endif
+ mov rtsin, [esp+4]
+%endif
PREROTATER r4, r3, r2, rtcos, rtsin
+%if mmsize == 8
+ mov r6, [esp]
+ movzx r6, word [rrevtab+r4*2]
+ mova [r1+r6*8], m0
+ inc r4
+ dec r3
+%else
%if ARCH_X86_64
movzx r5, word [rrevtab+r4-4]
movzx r6, word [rrevtab+r4-2]
@@ -929,6 +1055,7 @@ cglobal imdct_half, 3,12,8; FFTContext *s, FFTSample *output, const FFTSample *i
movhps [r1+r4*8], xmm1
%endif
sub r3, 4
+%endif
jns .pre
mov r5, r0
@@ -946,6 +1073,10 @@ cglobal imdct_half, 3,12,8; FFTContext *s, FFTSample *output, const FFTSample *i
%define rtsin r3
mov rtcos, [esp+8]
mov rtsin, [esp+4]
+%if mmsize == 8
+ sub rtcos, r0
+ sub rtsin, r0
+%endif
%endif
neg r0
mov r1, -mmsize
@@ -959,6 +1090,12 @@ cglobal imdct_half, 3,12,8; FFTContext *s, FFTSample *output, const FFTSample *i
DECL_IMDCT POSROTATESHUF
+INIT_MMX 3dnow
+DECL_IMDCT POSROTATESHUF_3DNOW
+
+INIT_MMX 3dnow2
+DECL_IMDCT POSROTATESHUF_3DNOW
+
INIT_YMM avx
%if HAVE_AVX
--
1.7.9.2
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment